Question 1090
A synthetic instruction-tuning dataset includes multiple near-identical variants of the same customer complaint, making the held-out set accidentally contain paraphrases of training examples. Which curation control improves generalization evidence?
Explanation
Correct Answer
Near-duplicate examples can contaminate evaluation even without exact text matches. Family-level partitioning keeps related supervision from leaking into held-out measurement.
Incorrect Answers
A longer context window does not restore independence between contaminated training and test samples.
Deleting difficult evaluations after scoring biases reported improvements instead of fixing leakage.
Independent row splitting can place essentially the same technical decision on both sides of a train-test boundary.